COMP6237 Data Mining Data Mining & Machine Learning with Big Data. Jonathon Hare

Transcription

1 COMP6237 Data Mining Data Mining & Machine Learning with Big Data Jonathon Hare

2 Contents Going to look at two case-studies looking at how we can make machine-learning algorithms work on big data data that is too large to be effectively stored on a single machine or processed in a time-efficient manner by a single machine. Specifically: K-Means clustering using Hadoop Stochastic Gradient Descent using Spark

3 Distributing K-Means with MapReduce

4 Recap: K-Means Clustering K-Means is a classic clustering algorithm for grouping data into K groups with each group represented by a centroid: The value of K is chosen K initial cluster centres are chosen Then the following process is performed iteratively until the centroids don t move between iterations (or max allowed iterations is reached): Each point is assigned to its closest centroid The centroid is recomputed as the mean of all the points assigned to it. If the centroid has no points assigned it is randomly re- initialised to a new point. The final clusters are created by assigning all points to their nearest centroid.

5 K-Means Demo

6 K-Means with Big Data Let s assume that we re building a visual image search system for millions of images (ala Google Goggles ). We d do this by decomposing an each image into a histogram of visual words that can be effectively indexed. A visual word vocabulary can be created by clustering SIFT features

7 K-Means with Big Data Typical image might have ~3000 SIFT Features Each one is a 128-dimensional vector Assuming double-precision fp (64bits/element) each image has ~3MB of features This kind of image search requires big vocabularies Say 1 million Assume we ve got 1 million images: We want to cluster 3B features into 1M clusters! Features take up ~3TB storage; centroids ~1GB.

8 How can K-Means be parallelised? Fundamentally the algorithm has two steps that repeat: Centroid assignment Centroid update These are independent, and both can be parallelised

9 Parallelised Assignment During assignment the centroids are static, so given enough processing resource each vector could be assigned in parallel. The process of searching for the closest centroid can also be parallelised by computing distance to each centroid in parallel. This looks a lot like a Map operation: Map(<null, vector>) - > <centroid, vector>

10 Parallelised Update Each centroid update is independent of the others, so they can also be performed in parallel. Computing the new centroid is a matter of accumulating the corresponding elements of all the assigned vectors, and dividing by the number of assigned vectors. This could (potentially) be parallelised on a per-element basis (although there might be practical limitations). This looks a lot like a Reduce operation: Reduce(<centroid, list(vector)>) - > <null, centroid>

11 MapReduce K-Means K-Means almost fits the MapReduce model Assign features to Centroids Recalculate Centroids Assign features to Centroids Recalculate Centroids Assign features to Centroids Recalculate Centroids... Assign features to Centroids Recalculate Centroids

12 MapReduce K-Means Caveats Map function needs to have access to the centroids in addition to its normal inputs Can use a fixed number of iterations but if we want to stop on convergence, the Reduce functions need to be able to communicate this

13 Basic Hadoop MapReduce K-Means Hadoop Mapper can load the current centroids from HDFS into memory during setup. Allows centroids to be used over all vectors belonging to the InputSplit Map function doesn t need to actually output the centroid vector an ID is fine (and much smaller) Unfortunately, you re still emitting every input vector! This is less of an issue because the cost of data transfer is likely to be outweighed by the cost of finding the closest centroid (assuming many centroids).

14 Basic Hadoop MapReduce K-Means The reducer doesn t actually care about the centroid vector (hence use of IDs in the map isn t a problem) just needs to know those vectors that have been assigned to a centroid in order to compute the new centroid vector For convergence checking, we could have a single Reducer that aggregates convergence of all reduces, or we could write convergence state to HDFS.

15 K-Means Initialisation To start K-Means, you need to have an initial set of K centroids. How can these be generated efficiently and effectively?

16 Random Generated Centroids In some cases, you can just generate some random vectors as a starting point. This only works well if the vectors you generate cover the space of the data vectors. In the case of SIFT features, the space, although bounded, is highly non-uniform and contains many areas with very low density, and others with very high density.

17 Random Sampled Centroids Choosing the centroids by randomly uniformly sampling the actual data is often a good starting point that works well in practice. How can we efficiently sample in our big data scenario? We don t necessarily know exactly how many features we have Finding out is relatively expensive because it means reading through all the data.

18 Hadoop MapReduce Sampling Synopsis: sample K items from an unknown amount of data Input: collection of files containing feature- vectors; read one vector at a time. Mapper: Map function: Input: <Key: null, Value: vector> Outputs: nothing Read the input vectors into the mapper memory. Output on shutdown: <Key: null, Value: vector> Output K uniformly randomly sampled vectors on shutdown of the mapper. Reduce function: Input: <Key: centroid_id, Value: [vectors of the items assigned to the cluster]> Outputs: <Key: Null, Value: centroid vector> Average the vectors to compute the new centroid, and emit this with a null key.

19 Canopy Clustering Fast algorithm to compute approximate clusters based on a pair of distance thresholds. Can be implemented in a single Hadoop MapReduce round: Each Mapper processes a subset of the data and uses the canopy algorithm to produce a set of clusters ( canopies ). Canopy clustering is applied to the centroids of all the mapper canopies (in a single Reducer) to produce the final set of initial centroids.

20 Better Hadoop MapReduce K-Means We ve dealt with initialisation and a basic Hadoop MapReduce K-Means implementation, but there are still two coupled problems: Finding the closest centroid (nearest-neighbour search) is slow. Every round of maps is emitting all the vectors to the reducers, resulting in massive amounts of data transfer.

21 Efficient Neighbour Search Rather than brute-force search (computing the distance of every point to every centroid), we can use an approximate method. for example, an ensemble of KD-Trees works well for SIFT-like vectors. Reduces computation from O(NK) to O(N log(k))

22 Reducing Data Transfer The Hadoop MapReduce framework has an additional concept called a Combiner. Combiners aim to reduce the data sent to the reducer by performing aggregation of data before it is sent to the reducer. A combiner behaves like a reducer that runs on the data output from mappers on a single machine. Because the combiner and mappers run on the same machine, there is no network I/O.

23 Reducing Data Transfer We can modify the MapReduce K-Means implementation to use a Combiner Output average vector and the number of samples used to produce it: Map(<null, vector>) - > <centroidid, <vector, 1>> Combine(<centroidID, list(<vector, 1>)>) - > <centroidid, <local_centroid_vector, count>> Reduce(<centroidID, list(<vector, count>)>) - > <null, centroid_vector>

24 Distributed Stochastic Gradient Descent with Spark

25 Gradient Descent Gradient Descent is a set of common first order optimisation algorithms. Attempts to find a (local) minimum of a differentiable function. Take steps proportional to the negative gradient of the function at the current point.

26 Machine Learning In machine learning we consider the problem of minimising an objective function that has the form of a sum: nx Q(w) = Q i (w) i=1 where w is the parameter vector that you re trying to estimate, and each summand function Qi is associated with the i-th observation of the training set.

27 Batch Gradient Descent (BGD) Each iteration of optimisation has the following form: w := w rq(w) =w P n i=1 rq i(w) learning rate The summation can be performed in parallel

28 Example: Linear Regression Problem: fit a (straight) line (y=w0x+w1) to some data, minimising the error between the points and the line. Objective function is: Q(w) = 1 N NX (y i (w 0 x i + w 1 )) 2 i=1 Or, in terms of the summand function: Q i (w) = 1 N (y i (w 0 x i + w 1 )) 2

29 Example: Linear Regression Differentiating w.r.t. w gives: Q i (w) = 2x i(y i (w 0 x i + w 1 )) w 0 N Q i (w) = 2(y i (w 0 x i + w 1 )) w 1 N So we now have a simple way to iteratively compute an estimate for w given some data

30 BGD Demo

31 Stochastic Gradient Descent (SGD) BGD becomes very expensive with large datasets Each iteration has to sum over all summand function gradients. SGD uses only a single observation to estimate the gradient at each step: w := w rq i (w) This scales better, although doesn t converge as smoothly Unfortunately, it s not easy to parallelise

32 SGD Demo

33 Mini-Batch SGD (MBSGD) MBSGD is a compromise between BGD and SGD It uses a small number of observations at each iteration The small sample is called a mini-batch The gradients of the summand for each observation in the mini-batches can be computed in parallel. Convergence is smoother than SGD

34 MBSGD Demo

35 Why is this useful? You ve used Gradient Descent before: The backpropagation algorithm! Directly applicable to many ML approaches: Training SVMs Training Neural Nets (inc. Deep Learning) Matrix factorisation Funk SVD used in Netflix

36 Big Data Gradient Descent Assume we have a big (distributed) dataset of observations and want to perform some kind of ML using gradient descent. Mini-batches could be formed making use of data locality to allow for data-local processing. But, processing of mini-batches is sequential: process batch -> update params -> process batch -> update params -> Not making effective use of computational resources

37 Downpour SGD Approach developed by google for training very large deep NNs using SGD Distributed processing based on a master-worker model. Master maintains current parameters, and updates them on receipt of gradients from a worker. Workers compute gradients of mini-batches

38 Downpour SGD Before starting to compute the gradients, each worker asks the master for the current parameters. Gradients are returned to the master as soon as the worker has finished, and the master immediately applies them to update the parameters (using Adagrad for learning rates). Obviously this means that many of the workers will be using out-of-date parameters. Strangely enough, this doesn t seem to matter!

39 Downpour Demo

40 Downpour SGD in Spark Implementation in Spark is trivial: Dataset represented by an RDD The master is the node on which you run the program; it sets up a server that provides methods to get the current parameters and update the parameters with new gradients. Communications could use anything: HTTP, RMI, even MPI The update method takes a gradient vector and uses it to alter the parameters of the model.

41 Downpour SGD in Spark The master then uses the mappartitions action with a function that: Fetches the current parameters from the master Computes the gradients of each mini-batch (the minibatch being defined as a RDD partition that is basically equivalent to a block of the underlying file on HDFS). Sends a message to the master informing it of the computed gradients.

42 Summary Machine learning with big data is challenging. Many considerations need to be taken into account. Minimising data transfer; optimising computation. Hadoop-MapReduce and Spark provide programming frameworks in which you can implement different types of algorithms.